I have previously been using a technique involving weighted means of ego corrections from recorded log files to tune VE tables.
Tonight I had a thought that this was not really the best way to do it, since the distribution of the samples is truncated at either end, and means don't really account for the spread of data. In other words, my weighted mean is a sample mean, not the mean of the underlying population.
So I plotted histograms of the ego correction of the closest matching samples to each table point. The central limit theorem seems to hold, so I get a collection of "mostly" Gaussian distributions - I can certainly fit Gaussians to the data - from which I can estimate population mean, standard deviation etc. Interestingly, some cells appear to have 2 populations, which implies there is some mode of operation I haven't properly accounted for...still, it's something worth noting because the weighted mean method won't pick this up, and will always give quite wrong answers for these type of cells, instead of there being a 50/50 chance.
So the question is which way is the best? Experimentally, they give similar, but subtly different numbers. (Not driven a car with a VE table tuned by the Gaussian fitting method yet, since I only invented it tonight!)
